Stanford's Matryoshka Attribution Tops Mechanistic Interpretability Benchmarks at 2.9x Runner-up

A Stanford team (including Aryaman, Chris Potts, and Dan Jurafsky) released a paper on 09-24 titled Matryoshka Attribution (MAttr), proposing a gradient-descent-based neural network attribution method that identifies which parts of a network are responsible for a given behavior. It dramatically sets a new SOTA on a mechanistic interpretability benchmark and is worth watching.

Confirmed

• Positioning: The authors argue against relying on axiomatic definitions of attribution results, instead framing attribution—a classic interpretability problem—as a training objective and directly optimizing via gradients for the target task; this is explicitly positioned as a response to Bilodeau, Jaques, Koh, and Kim's "Impossibility Theorem" paper.
• Technical approach: MAttr combines the strengths of causal interventions, gradient attribution, and mask learning, using a simple differentiable sigmoid top-k operator to parameterize causal interventions—no sparsity losses or straight-through tricks needed. Randomizing the top-k budget during training induces an attribution ranking.
• Benchmark results: On the mechanistic interpretability benchmark, its score is 2.9 times that of the runner-up; the authors report setting a new SOTA after roughly 500 training steps, using judge scores as the reward.
• Real-world case: MAttr is not limited to representation-level attribution—it can trace an LLM's qualitative behavior to the specific subset of parameters updated during training. Using StrongREJECT scores as the reward for reinforcement learning attribution and applying the mask to model weights between two checkpoints, the results show that Llama 3.1's refusal behavior is localized to 1% of the parameter differences.

Why it matters

• The method sidesteps the "impossibility theorem's" theoretical objections to axiomatic attribution, offering an optimizable engineering path for interpretability research.
• Being able to localize high-level behaviors (such as refusals) to a tiny fraction of parameters provides a practical tool for understanding and editing large model behavior (e.g., removing refusals).

2026-09-24 ~ 2026-09-24 · 7 related posts

• [source] Matryoshka Attribution tops interpretability benchmark by 2.9x, traces LLM refusals to 1% of weights — aryaman2020 · 2026-09-24
• Stanford's Matryoshka Attribution tops mech interp benchmark at 2.9x the runner-up — aryaman2020 · 2026-09-24
• MAttr technical details: differentiable sigmoid top-k merges causal interventions and mask learning — aryaman2020 · 2026-09-24
• MAttr turns attribution into a training objective, hitting MIB SOTA in just 500 steps — aryaman2020 · 2026-09-24
• [source] Attribution as a training objective: MAttr answers the Impossibility Theorems paper — aryaman2020 · 2026-09-24
• [source] MAttr traces Llama 3.1 refusals to 1% of parameter delta, resetting them disables refusal — aryaman2020 · 2026-09-24

1 near-duplicate retellings: ChrisGPotts